Description of LIM Decadal Forecasts
Near-realtime decadal forecasts are generated using the LIM described in Newman (2013). In the LIM, the dynamical evolution operator is estimated from the observed statistics of yearly-averaged global surface temperature variations over the years 1902-2009. (Also see Penland and Sardeshmukh 1995 for a complete treatment of LIM and its construction.)
The version of the LIM used in these web pages differs from the Newman (2013) LIM in two key details: surface land (2m) temperature data in the initialization is from the GHCN+CAMS (Fan and van den Dool 2008) dataset rather than the CRU TS 3.10 dataset, and anomalies are reported relative to the 1980-2009 climatology. The former change is made so that forecasts can be initialized in near real-time. All qualitative (and generally quantitative) aspects of the model remain the same.
Newman (2013) obtained gridded surface land temperatures from the CRU TS 3.10 dataset. However, since this dataset covered only the 1901-2009 period, we instead use the GHCN+CAMS, which does not cover the period prior to 1948. However, this dataset compares reasonably well to the CRU dataset over the common 1948-2009 period (see Fan and van den Dool 2008). We use a multiple linear regression between the CRU3.10 and GHCN+CAMS temperature anomalies, determined over the 1948-2009 period, to "predict" the initial CRU3.10 anomalies that are then used to initialize the LIM forecasts. (Thanks to Jon Eischeid for this suggestion.) This appears to have very little impact on skill of the hindcasts; in fact, land skill is more sensitive to which surface temperature dataset is used for verification.
Monthly anomalies are averaged with a 12-month running mean. All maps of forecast and verification anomalies are displayed relative to a 1980-2009 seasonally varying climatology. However, the climate indices are not recentered, nor are the hindcasts. Values of anomalies represent the average value within a 2deg latitude x 5deg longitude grid box.
Due to data availability limitations, forecast initializations generally are a few months behind.
The prediction model was validated using a jackknifing procedure in which 10% of the data are removed at a time to serve as independent data. This allows generation of hindcasts for the entire period of record. Maps of the resulting cross-validated hindcast skill for forecast leads of 2-5 years and 6-9 years are available in Newman (2013). All hindcasts for the post-1960 period (to match the period of the CMIP5 decadal hindcast experiment), along with their corresponding verifications, are in the hindcast archive. Forecasts initialized in 2009 and later are made using the linear dynamical operator determined from the entire 1902-2008 period.
All LIM forecasts on these webpages represent ensemble mean forecasts. Also, LIM not only produces forecasts but also allows for an a priori estimate of forecast skill. See Newman (2013) for more details. If time permits, such an estimate will be provided along with the forecasts on these pages.